Application Of Deep Learning In Nonlinear Process Fault Diagnosis

With the development of science and the advancement of technology,the industrial systems in modern large-scale production have become larger and larger,and the complexity of the system has also multiplied.Once an industrial system fails,people's property and personal safety are seriously threatened while causing waste of resources.Therefore,there is an urgent need to improve the reliability and safety of complex industrial systems,reduce accident risks,and improve economic efficiency.With the development of computers,sensors and communication technologies,modern industrial systems have shown a new trend towards large-scale and complicated development.The data reflecting the operating mechanism and state of the system presents large data volume,multi-modality,uncertainty and nonlinearity.With the "big data" feature,traditional data-driven fault diagnosis methods are not suitable.In recent years,deep learning has developed rapidly in academia and industry.It has significantly improved the accuracy in many traditional recognition tasks,and shows its superior feature extraction and ability to handle complex recognition tasks.So,a large number of experts and scholars take their attention to it.Its theory and application are being studied.Deep Belief Network(DBN)is a deep learning network,it is widely used in recent years.DBN not only has powerful automatic feature extraction capability but also it has advantages in processing high dimensionality and nonlinearity.It has achieved in the fields of image processing and speech recognition.In the context of data-driven fault diagnosis,this paper focuses on the problems and challenges faced by deep learning in fault diagnosis,and studies the application of DBN in fault diagnosis,as follows:1)Study the problems and challenges when deep leaning is used in fault diagnosis.Analyze the DBN,and point out its inadequacies in batch process fault diagnosis.2)For the nonlinear batch process with time correlation,a batch process fault diagnosis method based on DBN and long-short memory(LSTM)network is established.Firstly,the feature extraction ability of the DBN is used to extract features,and then the time correlation analysis is performed on the feature level.The DBN-LSTM can effectively detect various faults in a kind of semiconductor etching process simulation experiment,and has high fault identification accuracy,reaching more than 98%,which is 12% higher than the DBN and is about 6% higher than the CNN.3)Summarize this article and present ideas for future research work.